3102 IEICE TRANS. COMMUN., VOL.E92–B, NO.10 OCTOBER 2009 PAPER Resource Minimization Method Satisfying Delay Constraint for Replicating Large Contents Sho SHIMIZU †a) , Student Member, Hiroyuki ISHIKAWA †∗ , Nonmember, Yutaka ARAKAWA † , Naoaki YAMANAKA † , and Kosuke SHIBA †† , Members SUMMARY How to minimize the number of mirroring resources un- der a QoS constraint (resource minimization problem) is an important is- sue in content delivery networks. This paper proposes a novel approach that takes advantage of the parallelism of dynamically reconfigurable pro- cessors (DRPs) to solve the resource minimization problem, which is NP- hard. Our proposal obtains the optimal solution by running an exhaustive search algorithm suitable for DRP. Greedy algorithms, which have been widely studied for tackling the resource minimization problem, cannot al- ways obtain the optimal solution. The proposed method is implemented on an actual DRP and in experiments reduces the execution time by a factor of 40 compared to the conventional exhaustive search algorithm on a Pentium 4 (2.8 GHz). key words: content delivery network, replica placement, dynamically re- configurable processor, exhaustive search 1. Introduction Demand continues to grow for downloading rich contents, for example DVD-quality or high definition videos, through the Internet. Two factors are the key to meeting this demand: local content sources and adequate transfer capacity. Opti- cal networks can provide the high-speed and high-capacity pipes needed; they are now commonly used in backbone net- works and can handle bandwidth-consuming applications if the transfer distances are reasonable. This paper focuses on the other factor, an shows how to determine where to site content sources. Identifying the optimum number and location of con- tent sources (servers) involves an understanding the trade- offs between performance and cost. Using just a few servers is very effective in reducing initial investment costs but the servers will experience extremely high loads since they must deal with simultaneous download requests from many clients. Moreover, the average transfer distance is high which degrades the QoS and indeed overall network perfor- mance. The content delivery network (CDN) was proposed to improve network resource utilization efficiency for large contents distribution [1], [2]. CDN consists of two types of servers: origin server and replica server. The number of ori- Manuscript received February 10, 2009. Manuscript revised June 10, 2009. † The authors are with the Graduate School of Science for Open and Environmental Systems, Keio University, Yokohama-shi, 223- 8522 Japan. †† The author was with IPFlex Inc., Tokyo, 141-0021 Japan. ∗ Presently, with Kansai Electric Power Co. Inc. a) E-mail: shimizu@yamanaka.ics.keio.ac.jp DOI: 10.1587/transcom.E92.B.3102 gin servers is usually one (for each contents provider), and the many replica servers are spread throughout the service area. Origin server holds the original contents and delivers them to the replica servers as needed to ensure user requests can be satisfied. The contents stored in a replica server are called replicas. CDN promises high-speed downloads since the client downloads the data from the server nearest to the client in terms of network connectivity. In CDN, replica placement impacts the performance which includes the load on the origin server and the net- work since data placement decisions must be made on a per content basis and be made dynamically in response to user requests. Minimizing the number of mirroring resources (servers) under a Quality of Service (QoS) constraint is a key issue in CDN, so research in this area has been quite active. A tough problem to select which nodes should host which replicas. The distance between two nodes is used as a metric for QoS in CDN. A request must be resolved by a server within the distance specified by the request because all clients want to download contents within the allotted time period. Ev- ery node knows the nearest replica server that holds the re- quested data and the request is sent to the replica server that is closest to the client. The goal is to find a replica place- ment that satisfies all requests without violating any range constraint, and that minimizes the update and storage cost at the same time. This paper emphasizes the optimization of the number of replicas under the delay constraint. Replica placement problem is derived from the set cover problem which is known to be NP-hard [3]. There- fore, calculation time increases rapidly with network scale. Greedy algorithms have been widely studied since they yield sub-optimal solutions reasonably quickly [4]–[11]. How- ever, it has been proven mathematically that no greedy algo- rithm can attain the optimal solution [3]. Sub-optimum so- lutions have higher replicating cost, i.e. the number of repli- cas, than the optimal solution. The goal then is to secure the optimal replica placement within some practical time. Our solution to obtaining the optimal solution to the replica placement problem is based on combining advanced processors with suitable algorithms. It is not realistic to ob- tain the optimal solution with a Neumann-type processor given the number of all solution candidates. To drastically reduce the calculation time, we propose a novel approach that uses an exhaustive search algorithm that suits the paral- lelism offered by a dynamically reconfigurable parallel pro- Copyright c  2009 The Institute of Electronics, Information and Communication Engineers